AI Research
Understanding
the universe.
We are advancing the scientific frontier of Artificial General Intelligence. Our research focuses on capabilities, alignment, and robust reasoning.
Publications
Recent papers, architectural breakthroughs, and evaluations.
April 2026
Scaling Laws for Mixture-of-Experts
An empirical study demonstrating optimal routing strategies and parameter distribution when scaling MoE architectures beyond 1T parameters.
ArchitectureScaling
January 2026
Grokking and Latent Reasoning
Investigating the sudden emergence of structured logic and reasoning capabilities in sparsely activated transformer networks.
InterpretabilityMath
Safety & Alignment
Building powerful models requires an equal commitment to steering them. Our alignment team focuses on adversarial robustness, mechanistic interpretability, and truthful generation.
- Red Teaming methodologies for zero-day vulnerabilities
- Constitutional AI constraints mapping
- Reward modeling for complex, multi-turn reasoning tasks